I=6 Zuckerberg said they are already seeing signs of self-improvement in the models, bringing superintelligence within reach
Zuckerberg said they are starting to see glimpses of self-improvement in the models, giving way for superintelligence.
" I think the most exciting thing this year is that we’re starting to see early glimpses of self-improvement with the models, which means that developing super intelligence is now in sight," he said (min 1:17).
He said they don’t know when superintelligence will be ready but it’s better to act as if it would be ready in the next two to three years.
“I just think that we should bet and act as if it’s going to be ready in the next two to three years…And if that is what you believe, you are going to invest hundreds of billions for whatever it takes to build out the strongest team.” (min 2:29).
He said Meta is focused on building personal superintelligence which people could use in their everyday lives such as managing their personal relationships and entertainment, which is different from that of other companies who want to automate all of the economic productive work in the society (min 5:02).
Zuckerberg pointed out that they are already seeing glimpses of the models (Llama) improving Facebook algorithm and other AI systems across the company autonomously (min 9:08).
I=5 Meta hires Mark Lee and Tom Gunter, Apple’s top AI researchers, after poaching their former boss, Bloomberg reported citing people familiar with the matter.
I=5 Meta plans to restructure its AI group for the fourth time in six months, the Information reported
Meta plans to restructure its artificial intelligence group for the fourth time in six months, the Information reported citing people familiar with the matter.
Meta’s Superintelligence Labs will be divided into four; the TBD (“to be determined”) Lab, a products team, an infrastructure team, and Meta’s Fundamental AI Research lab.
The planned changes have not been announced internally and could still change.
The AGI foundations group, which focused on generative AI products and research is being dissolved and its leaders Ahmad Al-Dahle and Amir Frenkel are now working on "strategic MSL initiatives” and reporting to Wang.
Aparna Ramani, a Vice President, who has been leading Meta’s AI, data and developer infrastructure units, will lead MSL Infra.
Robert Fergus, who rejoined Meta in May will continue leading FAIR.
The Information reported yesterday on the planned changes.
I think it’s normal to see some turnover after a major restructuring. However, it’s definitely something to watch, especially if the turnover is occurring among the new hires.
I=5 Meta aims to release the next generation AI model, Llama 4.X or Llama 4.5 by year end
Meta’s Superintelligence team aims to release the next generation AI model, Llama 4.X or Llama 4.5 by year end, Business Insider reported citing two sources.
Meta’s earlier Llama 4 models, Scout and Maverick, were underwhelming.
The team behind Llama 4.X is now attempting to fix bugs and revive Llama 4.
I=6
Meta may use Google or OpenAI ai models until its own catch up, The Information reports
Meta’s Superintelligence team have considered integrating Google’s Gemini model or OpenAI’s models with Meta AI, The Information reported citing people familiar with the conversations..
The idea is to use Google’s or OpenAI’s models until Meta’s own models advance.
Superintelligence’s key priority is to ensure Llama 5 is competitive, the report added.
Internally, Meta has already integrated external AI models such as Anthropic models into some internal tools for staff.
“We are taking an all-of-the-above approach to building the best AI products; and that includes building world-leading models ourselves, partnering with companies, as well as open sourcing technology,” a Meta spokesperson said.
I=10
Meta will spend at least $600 billion through 2028 in USA to power AI, Zuckerberg said
Mark Zuckerberg said at the White House Dinner that they will spend at least $600 billion through 2028 in USA to power AI (min 1:28).
Zuckerberg: I think um you know, all of the companies here are building just making huge investments in in the country in order to build out data centers and and infrastructure to uh power the next wave of innovation…
Trump: How much are you spending uh would you say over the next few years?
Zuckerberg: Oh gosh. Um I mean I think it’s probably going to be something like I don’t know at least $600 billion through 28 in the US.
Apple also said it will invest $600 billion in USA over four years- the investment going to suppliers across all 50 states, direct employment, investments in R&D, silicon engineering, AI, data centers, and infrastructure projects like Apple TV+ productions.
Assessment
CapEx of $600 billion over four years would imply $150 billion per year, which is well above current expectations. Some analysts already anticipate Meta’s CapEx reaching $100 billion in 2026. There are no further details from Meta or analyst commentary yet.
My view (confidence level: 60) is that Zuckerberg may be using a definition similar to Apple’s, so that Meta’s U.S. investment doesn’t appear weak in Trump’s eyes.
Using Apple’s definition, R&D expenses (my estimate (valuation model): $60 billion) and CapEx expectations for 2026 would already be around $150 billion.
Just saw that Zuckerberg immediately told Trump that he wasn’t sure about the exact number after mentioning the $600billion.
That’s interesting because it shows how easily companies can walk back from these figures and an announcement like that is not necessarily always planned
I=8
Meta CFO clarifies Zuckerberg’s $600b comment as total U.S. investment envelope through 2028
CFO Susan Li said at the Goldman Sachs 2025 Communacopia and Technology Conference that the $600 billion that Zuckerberg gave reflects their total envelope of their US. investment through 2028 (around min 29:00).
”So just to clarify, Mark’s comments are referring to sort of the total envelope of our planned U.S. investment from this year, so including 2025 through 2028,” she said. “So that includes obviously all the data center infrastructure that we are building in the states, but it also just includes, like all the investments that go towards supporting our U.S. Business operations, all the people we hire in the U.S.”
She said the TBD Labs inside Meta’s Superintelligence team is working on the next frontier models that will arrive in the next one or two years (min 23:37)
“We’re very excited about what we call TBD Labs. It’s a placeholder name that’s sort of stuck and feels actually very appropriate, in the sense that I think a lot of what the team is going to build is yet to be precisely shaped or determined. It’s a pretty small, few dozen people, very talent-dense set of folks. They are kind of working on the next generation of foundation models, and we hope that those will be at the frontier over the course of the next year or two.”
I=5
Meta recruits Yang Song as Superintelligence research principal, reporting to Shengjia Zhao
Meta has recruited OpenAI’s Yang Song as Superintelligence research principal, reporting to Shengjia Zhao (Chief Scientist), the Wired reported.
His resume seems impressive with Zuckerberg touting him as the cocreator of ChatGPT, GPT-4, all mini models, 4.1, and o3 at OpenAI in an internal memo.
I=4 Meta in talks with Google to use Gemini to improve its ad business, the Information reported
Meta and Google are discussing using Gemini to improve Meta’s ad business, the Informationreported citing people familiar with the discussions.
Sources said the talks are still in early stages.
A Meta spokesperson denied the report.
“We’ve always built our own industry-leading, proprietary ad targeting and recommendation systems,” the Meta spokesperson said in an email. “Separate from that, we regularly evaluate third-party tools for the purpose of benchmarking, which is the case here.”
Meta’s CFO recently said they see strong opportunity to use LLMs to improve its ad retrieval and recommendation system.
I=3 Meta is said to acquire chips startup Rivos amid reports that its own team is not progressing as quickly as Zuckerberg would like
Meta is acquiring chips startup Rivos, Bloomberg reported citing people familiar with the discussions.
Terms of the deal were not disclosed but Rivos was seeking a new funding at a $2 billion valuation in August.
Rivos is developing its own graphics processing unit (GPU) that powers AI.
Meta has its own team known as Meta Training and Inference Accelerator, that is seeking to develop its own chips but sources said its efforts aren’t progressing as quickly as Zuckerberg would like.
I=5 Meta Superintelligence Labs ditches internal engineering tools for being too slow, Business Insider reported
Meta Superintelligence Labs has ditched internal engineering tools for being too slow and is now using GitHub and Vercel, Business Insider reported citing internal memos.
Meta is also developing its own internal tool called Nest, which will eventually become a “de facto option”.
